Natarizki / CMLM-0.8B

CMLM-0.8B

C oding + M ath L anguage M odel — a Qwen3.5-0.8B fine-tuned for code generation and mathematical reasoning.

Model Details

Training Data

Training Hyperparameters

learning_rate: 2e-4
max_steps: 500
per_device_train_batch_size: 2
gradient_accumulation_steps: 8
effective_batch_size: 16
warmup_steps: 100
optimizer: adamw_8bit
gradient_checkpointing: unsloth
lora_r: 16
lora_alpha: 32
lora_dropout: 0
target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
packing: true
max_seq_length: 2048

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Natarizki/CMLM-0.8B", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Natarizki/CMLM-0.8B")

messages = [{"role": "user", "content": "Write a Python function to check if a number is prime."}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)

outputs = model.generate(inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Benchmarks

Note: CMLM generates longer, more detailed responses for coding tasks (hence lower tok/s but higher quality). Math and general domains show near-parity with base model throughput. Benchmarked on NVIDIA T4 with float32 inference via Unsloth.

Limitations

• Trained on 65K samples; may underperform on niche domains
• 2048 token context limit; not suitable for long-document tasks
• float32 training preserves accuracy but increases inference memory vs. quantized variants
• No vision capabilities despite Qwen3.5's native multimodal architecture

License

Apache 2.0 (inherits from Qwen3.5)

Acknowledgments

• Qwen Team for the base model
• Unsloth for efficient T4-compatible training
• Dataset authors: Magicoder, MetaMath, NuminaMath teams